Rare immune fixation electrophoretogram recognition system based on two-stage classification normal form

Abstract

The invention provides a rare immune fixation electrophoretogram identification system based on a two-stage classification normal form, which is not limited to classification of common disease types, but adds identification of rare disease types on the premise of ensuring common disease type identification precision, so that the identification accuracy of the common disease types is improved. The vacancy of an existing immunofixation electrophoretogram recognition method based on deep learning in the aspect of rare disease type recognition is made up, and universality and practicability are greatly enhanced; meanwhile, a double-stage classification normal form method is adopted, so that the class imbalance ratio of the data set is effectively improved, and the influence caused by class imbalance of the data set is reduced; finally, a loss function LMF specially used for solving data class imbalance in the medical field is adopted, the recognition precision of a few classes is improved on the premise that the recognition precision of most classes is guaranteed, and then the overall recognition precision is improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical testing and deep learning technology, and in particular relates to a rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm. Background Art

[0002] Multiple myeloma (MM) is a malignant plasma cell disease characterized by the abnormal proliferation of monoclonal immunoglobulins, leading to symptoms such as bone destruction, anemia, and renal insufficiency. In recent years, with the continuous development of medical diagnostic technology, immunofixation electrophoresis (IFE) has become an important diagnostic tool for monoclonal immunoglobulin-related diseases such as multiple myeloma. Immunofixation electrophoresis separates proteins by electrophoresis, uses specific antibodies to bind to the target protein, and stains and develops it. Monoclonal proteins appear as single concentrated bands, while polyclonal proteins appear as diffuse bands. It is used to diagnose diseases such as multiple myeloma. Based on the results of immunofixation electrophoresis, multiple myeloma can be divided into ten disease types, including six common disease types: IgA-k, IgA-λ, IgG-k, IgG-λ, k, λ, and four rare disease types: IgM-k, IgM-λ, biclonal, and oligoclonal.

[0003] In traditional immunofixation electrophoresis pattern recognition, clinicians need to spend a lot of effort manually checking IFE data, and the diagnostic process is highly subjective and labor-intensive. Deep learning, as one of the core technologies of artificial intelligence, has made significant progress in the field of medical image analysis in recent years. Its core advantage lies in its ability to automatically learn features from large amounts of data and apply them to tasks such as classification, detection, and segmentation. Deep learning-based immunofixation electrophoresis pattern recognition uses deep learning models such as convolutional neural networks (CNN) to automatically classify IFE patterns and identify the type of monoclonal immunoglobulins, providing important technical support for improving diagnostic efficiency and accuracy.

[0004] On the one hand, to address the high cost and subjectivity of traditional immunofixation electrophoresis pattern recognition, many researchers have introduced deep learning technology to achieve intelligent immunofixation electrophoresis pattern recognition based on deep learning. Although these methods have promoted the development of immunofixation electrophoresis pattern recognition to a certain extent, their versatility and practicality are still significantly limited because these studies are limited to the recognition of common disease types and ignore the existence of rare disease types.

[0005] On the other hand, many existing methods suffer from the class imbalance problem of immunofixation electrophoresis pattern datasets. This causes the model to tend to learn the features of the majority class while ignoring the minority class, resulting in poor recognition ability for the minority class, thus affecting the accuracy and practicality of deep learning-based immunofixation electrophoresis pattern recognition. Summary of the Invention

[0006] To solve the above problems, the present invention provides a rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm. The two-stage classification strategy solves the problem of data set class imbalance, and adopts a loss function specially designed for solving data class imbalance in the medical field. While ensuring the recognition accuracy of common disease types, the recognition accuracy of rare disease types is improved.

[0007] A rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm, including a first-stage classification paradigm network N1 and a second-stage classification paradigm network N2;

[0008] The first-stage classification paradigm network N1 is used to receive and determine whether the pathological type represented by the immunofixation electrophoresis pattern to be tested belongs to a rare disease type or a common disease type;

[0009] The second-stage classification paradigm network N2 is used to receive and determine the type of rare disease to which the immunofixation electrophoresis pattern to be tested belongs when the judgment result given by the first-stage classification paradigm network N1 is a rare disease type.

[0010] Furthermore, a rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm also includes a type classification module;

[0011] The type classification module is used to set a first actual type label and a second actual type label for the immunofixation electrophoresis patterns in the training set, thereby classifying each immunofixation electrophoresis pattern into common disease types and rare disease types, thereby obtaining common disease type subsets and rare disease type subsets; wherein the first actual type label is a specific classification of the common disease type and the rare disease type, and the second actual type label is a specific classification of the rare disease type;

[0012] The training method of the first stage classification paradigm network N1 is:

[0013] Each immunofixation electrophoresis pattern in the training set is used as the input of the first-stage classification paradigm network N1, and the judgment results of each immunofixation electrophoresis pattern output by the first-stage classification paradigm network N1 are obtained;

[0014] A first loss function is constructed by combining the judgment results of each immunofixation electrophoresis pattern output by the first-stage classification paradigm network N1 with the first actual type label of each immunofixation electrophoresis pattern, and the parameters of the first-stage classification paradigm network N1 are adjusted according to the first loss function until the first loss function is less than a set value, thereby obtaining a trained first-stage classification paradigm network N1;

[0015] The training method of the second stage classification paradigm network N2 is:

[0016] The immunofixation electrophoresis patterns that are judged as rare disease types by the first-stage classification paradigm network N1 are used as inputs to the second-stage classification paradigm network N2, and the judgment results of the immunofixation electrophoresis patterns are output by the second-stage classification paradigm network N2;

[0017] A second loss function is constructed by combining the judgment results of each immunofixation electrophoresis pattern output by the second-stage classification paradigm network N2 with the second actual type label of each immunofixation electrophoresis pattern, and the parameters of the second-stage classification paradigm network N2 are adjusted according to the second loss function until the second loss function is less than the set value, thereby obtaining a trained second-stage classification paradigm network N2.

[0018] Furthermore, common disease types include IgA-k type multiple myeloma disease, IgA-λ type multiple myeloma disease, IgG-k type multiple myeloma disease, IgG-λ type multiple myeloma disease, k type multiple myeloma disease, and λ type multiple myeloma disease; rare disease types include IgM-k type multiple myeloma disease, IgM-λ type multiple myeloma disease, biclonal multiple myeloma disease, and oligoclonal multiple myeloma disease.

[0019] Furthermore, the first loss function L cross for:

[0020]

[0021] Where N is the total number of immunofixation electrophoresis patterns in the training set; C is the total number of types in the classification task; y i,c is the one-hot encoding of the first actual type label of the i-th immunofixation electrophoresis pattern in category c; p i,c The probability value of the i-th immunofixation electrophoresis pattern belonging to category c predicted by the first-stage classification paradigm network N1.

[0022] Furthermore, the second loss function L LMF for:

[0023]

[0024] Where j∈{1,...,k} is the index of the specific classification of the rare disease type, k represents the total number of specific classifications of the rare disease type; y is the second actual type label, z y is the output logit of the second-stage classification paradigm network N2 for the second actual type label y; u is the logit value after interval adjustment for the second actual type label y; α and β are two adjustable hyperparameters; p t =softmax(z y ), is the predicted probability of the second stage classification paradigm network N2 for the second actual type label y; α t is the set category balance coefficient; γ is the set focusing parameter.

[0025] Furthermore, a rare class immunofixation electrophoresis pattern identification system based on a two-stage classification paradigm also includes a pre-processing module;

[0026] The preprocessing module is used to perform a preprocessing operation on each immunofixation electrophoresis pattern before inputting it into the first-stage classification paradigm network N1 or the second-stage classification paradigm network N2, wherein the preprocessing operation includes removing messy borders, retaining the electrophoresis channel area, and brightening, darkening, blurring, and adding noise to the immunofixation electrophoresis pattern.

[0027] Furthermore, the first-stage classification paradigm network N1 is a convolutional neural network mobilenetv2.

[0028] Furthermore, the second-stage classification paradigm network N2 is a convolutional neural network efficientnetv2.

[0029] Beneficial effects:

[0030] 1. The present invention provides a rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm. It is not limited to classifying only common disease types. Instead, it adds the recognition of rare disease types while ensuring the recognition accuracy of common disease types. It fills the gap in the existing deep learning-based immunofixation electrophoresis pattern recognition method in the recognition of rare disease types, and greatly enhances its versatility and practicality.

[0031] 2. The present invention provides a rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm. The two-stage classification paradigm method effectively improves the class imbalance ratio of the data set and reduces the impact of class imbalance in the data set. At the same time, the present invention adopts the loss function LMF specially used in the medical field to solve data class imbalance. While ensuring the recognition accuracy of the majority class, it improves the recognition accuracy of the minority class, thereby improving the overall recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A block diagram of the principle of a rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm provided by the present invention;

[0033] Figure 2 A flowchart of a rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm provided by the present invention;

[0034] Figure 3 A schematic diagram of the classification of a rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm provided by the present invention;

[0035] Figure 4 This is an example of the immunofixation electrophoresis pattern of IgA-k type multiple myeloma disease provided by the present invention;

[0036] Figure 5 This is an example of the immunofixation electrophoresis pattern of IgA-λ type multiple myeloma disease provided by the present invention;

[0037] Figure 6 This is an example of the immunofixation electrophoresis pattern of IgG-k type multiple myeloma disease provided by the present invention;

[0038] Figure 7 This is an example of the immunofixation electrophoresis pattern of IgG-λ type multiple myeloma disease provided by the present invention;

[0039] Figure 8 This is an example diagram of the immunofixation electrophoresis pattern of type K multiple myeloma provided by the present invention;

[0040] Figure 9 This is an example of the immunofixation electrophoresis pattern of λ-type multiple myeloma provided by the present invention;

[0041] Figure 10 This is an example diagram of the immunofixation electrophoresis pattern of IgM-k type multiple myeloma disease provided by the present invention;

[0042] Figure 11 This is an example diagram of the immunofixation electrophoresis pattern of IgM-λ type multiple myeloma disease provided by the present invention;

[0043] Figure 12 This is an example of the immunofixation electrophoresis pattern of oligoclonal multiple myeloma provided by the present invention;

[0044] Figure 13 This is an example diagram of the immunofixation electrophoresis pattern of biclonal multiple myeloma provided by the present invention. DETAILED DESCRIPTION

[0045] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0046] like Figure 1 As shown, a rare class immunofixation electrophoresis (IFE) pattern recognition system based on a two-stage classification paradigm includes a first-stage classification paradigm network N1, a second-stage classification paradigm network N2, a preprocessing module, and a type classification module;

[0047] The preprocessing module is used to perform a preprocessing operation on each immunofixation electrophoresis pattern before inputting it into the first-stage classification paradigm network N1 or the second-stage classification paradigm network N2, wherein the preprocessing operation includes removing cluttered borders, retaining the electrophoresis channel area, and brightening, darkening, blurring, and adding noise to the immunofixation electrophoresis pattern to perform data enhancement;

[0048] The first-stage classification paradigm network N1 is used to receive and judge whether the pathological type represented by the immunofixation electrophoresis pattern to be tested belongs to a rare disease type or a common disease type; wherein, common disease types include IgA-k type multiple myeloma disease, IgA-λ type multiple myeloma disease, IgG-k type multiple myeloma disease, IgG-λ type multiple myeloma disease, k type multiple myeloma disease, λ type multiple myeloma disease; rare disease types include IgM-k type multiple myeloma disease, IgM-λ type multiple myeloma disease, biclonal multiple myeloma disease, oligoclonal multiple myeloma disease; the immunofixation electrophoresis patterns of six common disease types IgA-k, IgA-λ, IgG-k, IgG-λ, k, λ and four rare disease types IgM-k, IgM-λ, oligoclonal, biclonal are as follows Figures 4 to 13 shown.

[0049] The second-stage classification paradigm network N2 is used to receive and determine the type of rare disease to which the immunofixation electrophoresis pattern to be tested belongs when the judgment result given by the first-stage classification paradigm network N1 is a rare disease type;

[0050] The type classification module is used to set a first actual type label and a second actual type label for the immunofixation electrophoresis patterns in the training set, thereby classifying each immunofixation electrophoresis pattern into common disease types and rare disease types, thereby obtaining common disease type subsets and rare disease type subsets; wherein the first actual type label is a specific classification of the common disease type and the rare disease type, and the second actual type label is a specific classification of the rare disease type;

[0051] The training method of the first stage classification paradigm network N1 is:

[0052] Each immunofixation electrophoresis pattern in the training set is used as the input of the first-stage classification paradigm network N1, and the judgment results of each immunofixation electrophoresis pattern output by the first-stage classification paradigm network N1 are obtained;

[0053] The judgment results of each immunofixation electrophoresis pattern output by the first stage classification paradigm network N1 and the first actual type label of each immunofixation electrophoresis pattern are used to construct a first loss function, and the parameters of the first stage classification paradigm network N1 are adjusted according to the first loss function until the first loss function is less than the set value, thereby obtaining a trained first stage classification paradigm network N1; specifically, the first loss function L cross for:

[0054]

[0055] Where N is the total number of immunofixation electrophoresis patterns in the training set; C is the total number of types in the classification task; y i,c is the one-hot encoding of the first actual type label of the i-th immunofixation electrophoresis pattern in category c; p i,c The probability value of the i-th immunofixation electrophoresis pattern belonging to category c predicted by the first-stage classification paradigm network N1.

[0056] Specifically, in the first stage of training the classification paradigm network, step N1, the six calibrated common disease types (IgA-k, IgA-λ, IgG-k, IgG-λ, k, and λ) and one rare disease type are used as input. The training process is the same as that of a typical neural network, starting with forward inference and loss calculation, followed by backpropagation to optimize network parameters. The goal is to obtain a model capable of classifying the six common disease types (IgA-k, IgA-λ, IgG-k, IgG-λ, k, and λ) and one rare disease type.

[0057] The training method of the second stage classification paradigm network N2 is:

[0058] The immunofixation electrophoresis patterns that are judged as rare disease types by the first-stage classification paradigm network N1 are used as inputs to the second-stage classification paradigm network N2, and the judgment results of the immunofixation electrophoresis patterns are output by the second-stage classification paradigm network N2;

[0059] The second loss function is constructed by combining the judgment results of each immunofixation electrophoresis pattern output by the second stage classification paradigm network N2 with the second actual type label of each immunofixation electrophoresis pattern, and the parameters of the second stage classification paradigm network N2 are adjusted according to the second loss function until the second loss function is less than the set value, thereby obtaining the trained second stage classification paradigm network N2. Specifically, the second loss function L LMF for:

[0060]

[0061] Where j∈{1,...,k} is the index of the specific classification of the rare disease type, k represents the total number of specific classifications of the rare disease type; y is the second actual type label, z y is the output logit of the second-stage classification paradigm network N2 for the second actual type label y; u is the interval-adjusted logit value of the correct category y; α and β are two adjustable hyperparameters; p t =softmax(z y ), is the predicted probability of the second stage classification paradigm network N2 for the second actual type label y; α t is the set category balance coefficient; γ is the set focusing parameter.

[0062] Specifically, in the training step for the second-stage classification paradigm network N2, samples classified as rare disease types by the first-stage classification paradigm network N1 are used as input and trained using the LMF loss function, a specialized loss function in the medical field designed to address class imbalance in datasets. The training process is identical to that for the first-stage classification paradigm network N1, aiming to produce a model capable of detailed classification of the four rare disease types.

[0063] In the step of summarizing and aggregating the results of the two-stage classification paradigm network, the classification results of the first and second-stage classification paradigm networks N1 and N2 are summarized and merged to obtain the final classification results of the immunofixation electrophoresis map dataset of ten disease types.

[0064] Thus, the present invention provides a rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm, which mainly obtains an immunofixation electrophoresis pattern data set, pre-processes the immunofixation electrophoresis pattern data, re-classifies the immunofixation electrophoresis pattern data set, divides it into training set, test set and validation set, trains the first-stage classification paradigm network N1, trains the second-stage classification paradigm network N2, and summarizes and aggregates the results of the two-stage classification paradigm network. In the step of dividing the training set, test set and validation set, the immunofixation electrophoresis pattern data set is divided into a training set, a test set and a validation set according to a set random seed, and the division standard is that the training set, the test set and the validation set are independent and identically distributed and are all subsets of the data set.

[0065] Further, Figure 2 and Figure 3 The flowchart and classification diagram of the rare class immunofixation electrophoresis pattern recognition system based on the dual-stage classification paradigm according to an embodiment of the present invention are respectively shown. Figure 2 In step S1, the immunofixation electrophoresis pattern datasets of ten calibrated disease types are first obtained, including six common disease types: IgA-k, IgA-λ, IgG-k, IgG-λ, k, λ, and four rare disease types: IgM-k, IgM-λ, biclonal, and oligoclonal.

[0066] In step S2, the obtained immunofixation electrophoresis map samples of the ten disease types are preprocessed respectively, including removing the messy borders to retain the electrophoresis channel area, and brightening, darkening, blurring and adding noise to the samples for data enhancement.

[0067] In step S3, the categories of the pre-processed immunofixation electrophoresis data set are reclassified to improve the imbalance ratio of the immunofixation electrophoresis data set, thereby reducing the impact of the imbalance of the data set. The imbalance ratio formula is as follows:

[0068]

[0069] Among them, N min is the number of samples of the majority class (the class with the largest number of samples); N max is the number of samples of the minority class (the class with a small number of samples).

[0070] Specifically, the four rare disease types, IgM-κ, IgM-λ, biclonal, and oligoclonal, were each classified as one rare disease type, resulting in the immunofixation electrophoresis pattern dataset containing seven disease types: six common disease types (IgA-κ, IgA-λ, IgG-κ, IgG-λ, κ, λ), and one rare disease type. This approach improved the class imbalance ratio of the immunofixation electrophoresis pattern dataset from 0.0239 to 0.1714.

[0071] In step S4, the immunofixation electrophoresis pattern dataset is divided into a training set, a test set, and a validation set according to the set random seed. The division criteria are that the training set, the test set, and the validation set are independent and identically distributed and are all subsets of the dataset.

[0072] In step S5, the first stage classification paradigm network N1 is trained to obtain a model that can classify six common disease types IgA-k, IgA-λ, IgG-k, IgG-λ, k, λ and one rare disease type.

[0073] Specifically, as shown in the schematic diagram Figure 3As shown in the figure, forward propagation is first performed, inputting the reclassified six common disease types (IgA-k, IgA-λ, IgG-k, IgG-λ, k, λ) and one rare disease type into the first-stage classification paradigm network N1. The output is calculated through each layer of the network. Here, N1 uses the convolutional neural network mobilenetv2. The cross-entropy loss is then calculated based on the difference between the classification results and the dataset labels. Finally, backpropagation is performed based on the loss value to calculate the updated gradient information of the network N1 parameters until the network parameters are updated to the optimal level.

[0074] After the first stage classification results are obtained in step S5, in step S6, the second stage classification paradigm network N2 is trained to obtain a model that can classify the four rare disease types IgM-k, IgM-λ, biclonal, and oligoclonal in detail.

[0075] Specifically, as shown in the schematic diagram Figure 3 As shown in the figure, forward propagation is first performed, using samples classified as rare diseases in step S4 as input to the second-stage classification paradigm network N2. The network layers calculate the output. Here, N2 uses the convolutional neural network EfficientNetV2. Then, using the LMF loss function, specifically designed for addressing data class imbalance in the medical field, the loss value is calculated based on the difference between the classification result and the dataset label. Finally, backpropagation is performed based on the loss value to calculate the updated gradient information for the parameters of network N2 until the network parameters are optimized.

[0076] In step S7, the classification results obtained in step S4 and step S5 need to be summarized and combined to obtain the final classification results of the immunofixation electrophoresis pattern data set of ten disease types, including six common disease types IgA-k, IgA-λ, IgG-k, IgG-λ, k, λ and four rare disease types IgM-k, IgM-λ, biclonal, and oligoclonal.

[0077] The following is a table showing the results of classifying the collected immunofixation electrophoresis map samples of ten disease types using the first-stage classification paradigm network Mobilenetv2 and the second-stage classification paradigm network Efficientnetv2. Table 1 and Table 2 are the first-stage classification results and the second-stage classification results, respectively.

[0078] Table 1. Classification results of the first stage

[0079]

[0080] Table 2 Second stage classification results

[0081]

[0082] The two result tables above show that the average recognition accuracy for the six common disease types reached 97.62%, with an average F1 score of 0.98; the average recognition accuracy for the four rare disease types reached 98.51%, with an average F1 score of 0.99. This demonstrates that the rare class immunofixation electrophoresis pattern recognition system based on the dual-stage classification paradigm not only accurately classifies the six common disease types, but also achieves precise recognition of the four rare disease types.

[0083] In summary, existing deep learning-based immunofixation electrophoresis pattern recognition methods are limited to common disease types and ignore rare disease types, resulting in limited versatility and practicality. The rare immunofixation electrophoresis pattern recognition system of the present invention is not limited to classifying only common disease types. Instead, it incorporates the recognition of rare disease types while ensuring the accuracy of common disease type recognition. This fills the gap in rare disease type recognition of existing deep learning-based immunofixation electrophoresis pattern recognition methods and greatly enhances their versatility and practicality.

[0084] At the same time, existing deep learning-based immunofixation electrophoresis pattern recognition methods are mostly affected by the class imbalance problem of immunofixation electrophoresis pattern datasets, and have poor recognition capabilities for minority classes. The rare class immunofixation electrophoresis pattern recognition system of the present invention adopts a two-stage classification paradigm to effectively improve the class imbalance ratio of the dataset and reduce the impact of class imbalance. At the same time, this method adopts the LMF loss function specifically designed for solving data class imbalance in the medical field. While ensuring the recognition accuracy of the majority class, it improves the recognition accuracy of the minority class, thereby improving the overall recognition accuracy.

[0085] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may of course make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A rare class immunofixation electrophoresis pattern recognition system based on a two-stage classification paradigm, characterized by: Including the first stage classification paradigm network N1 and the second stage classification paradigm network N2; The first-stage classification paradigm network N1 is used to receive and determine whether the pathological type represented by the immunofixation electrophoresis pattern to be tested belongs to a rare disease type or a common disease type; The second-stage classification paradigm network N2 is used to receive and determine the type of rare disease to which the immunofixation electrophoresis pattern to be tested belongs when the judgment result given by the first-stage classification paradigm network N1 is a rare disease type.

2. The rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm according to claim 1, characterized in that: Also includes type classification module; The type classification module is used to set a first actual type label and a second actual type label for the immunofixation electrophoresis patterns in the training set, thereby classifying each immunofixation electrophoresis pattern into common disease types and rare disease types, thereby obtaining common disease type subsets and rare disease type subsets; wherein the first actual type label is a specific classification of the common disease type and the rare disease type, and the second actual type label is a specific classification of the rare disease type; The training method of the first stage classification paradigm network N1 is: Each immunofixation electrophoresis pattern in the training set is used as the input of the first-stage classification paradigm network N1, and the judgment results of each immunofixation electrophoresis pattern output by the first-stage classification paradigm network N1 are obtained; A first loss function is constructed by combining the judgment results of each immunofixation electrophoresis pattern output by the first-stage classification paradigm network N1 with the first actual type label of each immunofixation electrophoresis pattern, and the parameters of the first-stage classification paradigm network N1 are adjusted according to the first loss function until the first loss function is less than a set value, thereby obtaining a trained first-stage classification paradigm network N1; The training method of the second stage classification paradigm network N2 is: The immunofixation electrophoresis patterns that are judged as rare disease types by the first-stage classification paradigm network N1 are used as inputs to the second-stage classification paradigm network N2, and the judgment results of the immunofixation electrophoresis patterns are output by the second-stage classification paradigm network N2; A second loss function is constructed by combining the judgment results of each immunofixation electrophoresis pattern output by the second-stage classification paradigm network N2 with the second actual type label of each immunofixation electrophoresis pattern, and the parameters of the second-stage classification paradigm network N2 are adjusted according to the second loss function until the second loss function is less than the set value, thereby obtaining a trained second-stage classification paradigm network N2.

3. The rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm according to claim 2, characterized in that: Common disease types include IgA-k type multiple myeloma disease, IgA-λ type multiple myeloma disease, IgG-k type multiple myeloma disease, IgG-λ type multiple myeloma disease, k type multiple myeloma disease, and λ type multiple myeloma disease; rare disease types include IgM-k type multiple myeloma disease, IgM-λ type multiple myeloma disease, biclonal multiple myeloma disease, and oligoclonal multiple myeloma disease.

4. The rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm according to claim 2, characterized in that: The first loss function L cross for: Where N is the total number of immunofixation electrophoresis patterns in the training set; C is the total number of types in the classification task; y i,c is the one-hot encoding of the first actual type label of the i-th immunofixation electrophoresis pattern in category c; p i,c The probability value of the i-th immunofixation electrophoresis pattern belonging to category c predicted by the first-stage classification paradigm network N1.

5. The rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm according to claim 2, characterized in that: The second loss function L LMF for: Where j∈{1,...,k} is the index of the specific classification of the rare disease type, k represents the total number of specific classifications of the rare disease type; y is the second actual type label, z y is the output logit of the second-stage classification paradigm network N2 for the second actual type label y; u is the logit value after interval adjustment for the second actual type label y; α and β are two adjustable hyperparameters; p t =softmax(z y ), is the predicted probability of the second stage classification paradigm network N2 for the second actual type label y; α t is the set category balance coefficient; γ is the set focusing parameter.

6. The rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm according to claim 2, characterized in that: It also includes a pre-processing module; The preprocessing module is used to perform a preprocessing operation on each immunofixation electrophoresis pattern before inputting it into the first-stage classification paradigm network N1 or the second-stage classification paradigm network N2, wherein the preprocessing operation includes removing messy borders, retaining the electrophoresis channel area, and brightening, darkening, blurring, and adding noise to the immunofixation electrophoresis pattern.

7. The rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm according to claim 1, characterized in that: The first stage classification paradigm network N1 is a convolutional neural network mobilenetv2.

8. The rare class immunofixation electrophoresis pattern recognition system based on a dual-stage classification paradigm according to claim 1, characterized in that: The second-stage classification paradigm network N2 is a convolutional neural network EfficientNetV2.

Images